Mental health disorders affect hundreds of millions worldwide, yet professional counseling resources are severely limited. AI‑driven dialogue systems offer a scalable alternative, but current models suffer from two fundamental issues. First, they lack bidirectional understanding needed to capture the layered nature of emotional expression, especially during progressive disclosure where clients initially reveal surface symptoms and later expose deeper trauma; autoregressive (AR) models process inputs sequentially and cannot revise early interpretations when new evidence appears later. Second, they fail to incorporate the relational knowledge underlying clinical reasoning. To address these gaps, we introduce BiGraph‑Diffuse, the first large‑scale diffusion language model tailored for counseling, together with BiGraph‑RAG, a relation‑free graph‑structured retrieval strategy that relies solely on lightweight entity extraction and semantic linking. This design preserves inferential pathways from observable symptoms to potential underlying causes while incurring zero LLM token cost during indexing. The diffusion model supplies holistic bidirectional context, allowing the system to defer premature judgments during progressive disclosure, whereas the graph‑based retrieval captures structured clinical interconnections; the two components mutually reinforce each other. Extensive experiments demonstrate the effectiveness of BiGraph‑Diffuse, and we provide solid theoretical analysis supporting its design.
Review